maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity

[mlir][tosa] Fix dynamic shape inference in conv2d 

The comment in the code correctly states the equation for the shape inference as follows:

```
H = ((IH+pad_top+pad_bottom-(dilation_y*(KH-1)+1))/stride_y)+1
```

However the final operation is generated as `-` instead of `+`. I believe `+`
is indeed correct. For example if we have an image with dimension 6 and kernel
of dimension 3 (assuming padding is 0 and stride and dilation are both 1) we
are expecting 4 elements in the output (computed for image elements `(0, 1, 2)
x kernel`, `(1, 2, 3) x kernel`, `(2, 3, 4) x kernel` and `(3, 4, 5) x kernel`.
However currently only 2 elements are produced in the output.

Reviewed By: NatashaKnk

Differential Revision: https://reviews.llvm.org/D133208
This commit is contained in:
Anastasia Stulova 2022-09-06 15:07:29 -07:00 committed by Rob Suderman
parent 67dc6674c1
commit e28a5ceca9
2 changed files with 3 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
mlir/lib/Conversion/TosaToLinalg/TosaToLinalgNamed.cpp
View File

@ -93,7 +93,7 @@ getConvOutputDim(Location loc, Value initDim, Attribute padBeforeAttr,
Value subtract = builder.create<arith::SubIOp>(paddedAfter, addOne);
Value stride = reifyConstantDim(strideAttr, builder);
Value divide = builder.create<arith::DivUIOp>(subtract, stride);
return builder.create<arith::SubIOp>(divide, one);
return builder.create<arith::AddIOp>(divide, one);
}
// Creates a vector of the dynamic output dims for Conv2D and Depthwise_Conv2D

4
mlir/test/Conversion/TosaToLinalg/tosa-to-linalg-named.mlir
View File

@ -405,7 +405,7 @@ func.func @conv2d_dyn_w_h(%input: tensor<1x?x?x27xf32>, %weights: tensor<28x3x3x
// CHECK: %[[SUBTRACTED:.+]] = arith.subi %[[ADD_PAD_1]], %[[ADD_ONE]] : index
// CHECK: %[[STRIDE_H:.+]] = arith.constant 1 : index
// CHECK: %[[DIVIDED:.+]] = arith.divui %[[SUBTRACTED]], %[[STRIDE_H]] : index
// CHECK: %[[H_OUT:.+]] = arith.subi %[[DIVIDED]], %[[ONE]] : index
// CHECK: %[[H_OUT:.+]] = arith.addi %[[DIVIDED]], %[[ONE]] : index
// Computing output width
// CHECK: %[[C2:.+]] = arith.constant 2
@ -424,7 +424,7 @@ func.func @conv2d_dyn_w_h(%input: tensor<1x?x?x27xf32>, %weights: tensor<28x3x3x
// CHECK: %[[SUBTRACTED_0:.+]] = arith.subi %[[ADD_PAD_3]], %[[ADD_ONE_0]] : index
// CHECK: %[[STRIDE_W:.+]] = arith.constant 1 : index
// CHECK: %[[DIVIDED_0:.+]] = arith.divui %[[SUBTRACTED_0]], %[[STRIDE_W]] : index
// CHECK: %[[W_OUT:.+]] = arith.subi %[[DIVIDED_0]], %[[ONE_0]] : index
// CHECK: %[[W_OUT:.+]] = arith.addi %[[DIVIDED_0]], %[[ONE_0]] : index
// Running convolution
// CHECK: %[[PERM:.+]] = arith.constant dense<[1, 2, 3, 0]>